首页|基于有机光电突触晶体管的储池计算网络

基于有机光电突触晶体管的储池计算网络

扫码查看
原文链接
  • 万方数据
  • 维普
随着人工智能技术的快速发展,兼具高能效智能感知和计算的光电神经形态受到了广泛的关注.为了增强神经形态器件对不同应用场景的普适性,构建突触塑性能够灵活可调制的神经形态器件和电路至关重要.本文设计了一种基于非对称电极结构的有机突触晶体管,并利用液相制备的大面积有机超薄半导体作为光感知和计算材料.该光电突触晶体管表现出典型的光刺激后突触电流(excitatory postsynaptic potential,EPSC)、双脉冲易化(paired-pulse facilitation,PPF)和脉冲强度依赖塑 性(spike-amplitude-dependent plasticity,SADP),可以实现高能效的图像降噪预处理.为了进一步满足储池计算对于突触塑性可调性和网络非线性的要求,我们制备有机n型晶体管,设计了 一个基于p型和n型晶体管的突触模拟电路.该突触电路可实现从短程塑性(short-term synaptic plasti-city,STP)到长程塑性(long-term synaptic plasticity,LTP)的高度可调性,以及可配置的双脉冲易化特性,显著增强了突触晶体管的电流非线性度.基于有机突触模拟电路,我们构建了降噪预处理融合的储池计算系统,该储池神经网络对于去噪后的MNIST手写数据集表现出85％的识别准确率.此工作表明,有机光电突触晶体管有效的降噪功能和突触模拟电路的灵活调制能力对实现多模态、多尺度、高准确率和低功耗神经网络具有重要的意义.
Reservoir computing networks based on organic photoelectric synaptic transistors
With the rapid development of artificial intelligence technology,optoelectronic neuromorphic devices with energy-efficient intelligent sensing and computation have received widespread attention.In order to enhance the universality of neuromorphic devices for different application scenarios,it is crucial to construct neuromorphic devices and circuits whose synaptic plasticity can be flexibly modulated.In this paper,we design an organic synaptic transistor based on an asymmetric electrode structure and utilize a large-area organic ultrathin semiconductor prepared in the liquid phase as a material for photoreception and computation.The photoelectric synaptic transistor exhibits typical excitatory postsynaptic potential(EPSC),paired-pulse facilitation(PPF),and spike-amplitude-dependent plasticity(SADP),which can achieve energy-efficient image noise reduction.To further satisfy the requirements of reservoir computing for synaptic plasticity tunability and network nonlinearity,we prepare organic n-type transistors and design a synaptic analogue circuit based on p-type and n-type transistors.The synaptic circuit achieves a high degree of tunability from short-time plasticity to long-time plasticity,as well as a configurable PPF that significantly enhances the current nonlinearity of the synaptic transistor.Based on the organic synaptic analogue circuits,we further construct a noise-reducing preprocessing fusion reservoir computing system,which is a reservoir neural network that exhibits 85％recognition accuracy for denoised MNIST handwritten datasets.This work demonstrates that the effective noise reduction function of organic photoelectric synaptic transistors and the flexible modulation capability of synaptic analogue circuits are important for realizing multimodal,multiscale,high-accuracy and low-power reservoir computing.

organic synaptic transistorssynaptic analog circuitstunable synaptic plasticityPPFreservoir computing

杜晓松、顾健晖、彭立超、孙金瑞、姜赛、李昀

展开 >

常州大学微电子与控制工程学院,常州 213164

南京微结构国家实验室,人工微结构科学与技术协同创新中心,南京大学电子科学与工程学院,南京 210093

有机突触晶体管 突触模拟电路 高可调性 PPF 光电多模态储池计算

国家自然科学基金国家自然科学基金国家重点研发计划国家重点研发计划江苏省高层次双创人才计划常州市创新人才引进与培养领军项目常州市创新人才引进与培养领军项目常州市科技计划

62206030623740812022YFA12038022021YFA0715600JSSCRC2021534CQ20210116CQ20210084CJ20220085

2024

10.1360/SSC-2023-0239

中国科学(化学)
中国科学院

中国科学(化学)

CSTPCD北大核心
影响因子：0.685
ISSN：1674-7224
年,卷(期)：2024.54(4)
  • 30